SARCOVID: A Framework for Sarcasm Detection in Tweets Using Hybrid Transfer Learning Techniques

Balaji T K, Annushree Bablani, Sreeja S R, Hemant Misra

Abstract: The COVID-19 pandemic sparked a surge in online discussions, making sentiment analysis challenging due to the prevalence of sarcasm on social media. Identifying sarcastic expressions within the context of COVID-19 conversations poses a unique linguistic hurdle. To tackle this challenge, a novel framework called SARCOVID is proposed that leverages hierarchical transfer learning and ensemble techniques to detect sarcasm in the field. Through rigorous evaluation on a collected COVID-19 dataset, SARCOVID demonstrates superior performance in identifying sarcastic content with reduced bias compared to traditional methods. The findings reveal a significant presence of sarcasm in online COVID-19 discussions, underscoring the importance of robust sarcasm detection techniques. In a test, the framework outperforms other models with 0.61 accuracy on Sarcasm corpus V2. This approach not only advances sentiment analysis capabilities for evolving online conversations but also provides deeper insights into the nuanced expressions of sentiment on social media.